mburtsev/seafolk-danish-ais

Seafolk — Danish AIS aggregates

Danish waters counted, not tracked. The Danish Maritime Authority's open AIS archive is the only open European source that still contains the leisure fleet — Finland strips Class B for privacy, Norway excludes leisure craft under 45 m — and almost every pipeline throws that part away as noise. This dataset keeps it, as aggregates: how many distinct vessels of each kind were in a cell of sea in an hour (or a day), how many messages they sent, how fast the moving ones went, plus a per-day crossing count for every named Danish ferry line. Private boats — every Class B transponder and every pleasure craft whatever its transponder — appear only in cells shared by at least five of them. It is built by the Seafolk project and every number in it is reproducible from that repository.

Source and terms

Source: the Danish Maritime Authority (DMA) historical AIS archive, http://aisdata.ais.dk/, granted as open data on the basis of act no. 596 of 24 June 2005 on the further use of public sector information (the Danish PSI act).

The DMA's AIS data management policy (read verbatim 2026-09-12, see docs/DATA.md) states the conditions and the disclaimer:

The data recipient is responsible for handing the data received appropriately and in accordance with the law.

Data must not be combined with other data in a manner that will create data from which persons are identifiable, without a permit to do so from the Danish Data Protection Agency. In this connection, the data recipient should be considered the data responsible person in the sense of the act on personal data.

The Danish Maritime Authority does not guarantee the correctness of the AIS data transmitted and is not liable for any damages that may arise as a consequence of the use of AIS data, irrespective of how they have been caused. The Danish Maritime Authority is not liable in case of delay, interruption or loss of data due to failure of the transmission and communication facilities used or for other reasons beyond the control of the Danish Maritime Authority.

The Danish Maritime Authority can, at any time and without notice, change the nature of the data flow received, for example by changing the updating frequency, the area of coverage, safety updates, force majeure situations, etc.

The "persons are identifiable" condition is the one sentence that binds a derivative like this one, and the k ≥ 5 floor described under Privacy rule is how this dataset complies with it: no private vessel is distinguishable in any published cell, and nothing here is combined with a register that could name one.

Licence

The aggregates in this dataset are released under CC BY 4.0. Required attribution:

Data: Danish Maritime Authority (aisdata.ais.dk), aggregated by the Seafolk project

The DMA publishes no licence name and no attribution string of its own (see docs/DATA.md); CC BY 4.0 is the Seafolk project's choice for its own aggregates, permitted because the PSI act allows re-use and nothing in the DMA policy restricts derivatives.

Not for navigation. These are historical, aggregated counts with a known message-count bias and known coverage gaps. Nothing here is a position report, and nothing here may be used for navigation, collision avoidance, or any safety-of-life purpose.

Coverage

Bounding box: latitude 53–59 N, longitude 3–17 E, in hourly or daily UTC slices.

This box is not Denmark. It reaches into the German Baltic, the Sound, Kiel Bay and the Skagerrak. German-flagged Class B transponders outnumber Danish ones in every loaded year (0.32–0.36 of the fleet against 0.18–0.26). Read every number as "the fleet in this box", never as "the Danish fleet".

Loaded days — 2 122 in total, and the gaps are real gaps, not zeros:

Period	Days	Grain of the source file
2015-01-01 → 2015-12-31	365	monthly archives
2018-01-01 → 2018-12-31	365	monthly archives
2021-01-01 → 2021-12-31	365	monthly archives
2022-01, 2022-02	59	monthly archives (storms Malik, Nora)
2023-02, 2023-12	59	monthly archives (storms Otto, Pia)
2024-03-01 → 2026-08-26	909	daily archives, no gaps (2024: 306, 2025: 365, 2026: 238)

2016, 2017, 2019 and 2020 are not loaded at all. Any year-over-year series in this dataset therefore compares 2015 / 2018 / 2021 / 2024 / 2025 / 2026, and 2022 and 2023 contribute two months each.

Privacy rule

Private boats and the people on them are the whole reason this file exists. The private fleet is every Class B transponder AND every vessel whose ship_group is leisure, whatever transponder class it carries — a pleasure craft with a Class A radio is still a private boat, and 337–467 of them appear in the archive each year.

• Aggregates only for the private fleet. No MMSI, name, callsign, or track of a private vessel exists in any file here — no identifier column exists at all.
• k ≥ 5: every published leisure cell-day has at least five distinct vessels in it. Cells below the floor are dropped, not rounded or suppressed in place.
• The private hourly resolution-7 layer is never published, at any k. It does not survive the floor (whole-store measurement, 2026-09-12: at resolution 7 × hour the Class B layer would keep 7.1 % of the cell-hours and 20.8 % of the moving messages), and a single-vessel exact-distinct state is a membership oracle over MMSI. leisure_daily.parquet at resolution 5 × day is the private-fleet product, and it is the only place a leisure row is published.
• Ferries and commercial vessels are public and are named.
• The floor is enforced by a test that the export refuses to skip: scripts/test_export.py asserts that no private-fleet row has fewer than five vessels, that no leisure row appears in a Class A file, and that every file's columns and types are exactly the schema published below — an extra column is a failure, not a bonus. The export fails loudly.

Cost of the floor, measured 2026-09-12: 341 092 of 1 445 495 private-fleet cell-days survive — 23.6 % of the cells, but 82.0 % of the moving messages.

Files and schema

class_a_hourly_<year>.parquet — Class A, hourly, H3 resolution 7

One file per loaded year (2015, 2018, 2021, 2022, 2023, 2024, 2025, 2026). These files carry cargo, passenger, fishing and other only — a commercial vessel is public and required to broadcast, so no privacy floor is applied to it. Class A leisure is not here: it is part of the private fleet and lives in leisure_daily.parquet, under the floor.

Column	Type	Meaning
h3	uint64	H3 cell id, resolution 7 (~5 km²), geoToH3(lat, lon, 7)
hour	timestamp (UTC)	start of the hour
ship_group	string	cargo (Cargo + Tanker), passenger, fishing, other; resolved once per vessel-day, not per message. Never leisure
msgs	uint64	position/static messages counted in the cell-hour
moving_msgs	uint64	of those, messages with 0.5 kn < SOG < 100 kn
vessels	uint64	exact distinct vessel count
mean_sog	double	knots, mean over moving messages only; NULL when moving_msgs = 0

leisure_daily.parquet — the private fleet, daily, H3 resolution 5

Every Class B transponder or any vessel with ship_group = leisure whatever its class. The rule is applied per cell-hour, not per vessel: a boat that reports Class B one day and Class A another is under the floor on its Class B days in any case, and on its Class A days too when it is grouped as leisure.

Column	Type	Meaning
h3	uint64	H3 cell id, resolution 5 (~250 km²)
day	date (UTC)	calendar day
ship_group	string	same vocabulary, plus leisure = Sailing + Pleasure; a non-leisure group here is a Class B vessel of that group
msgs	uint64	messages in the cell-day
moving_msgs	uint64	of those, moving
vessels	uint64	exact distinct vessel count, always ≥ 5
mean_sog	double	knots, moving messages only; NULL when moving_msgs = 0

ferry_daily.parquet — named ferry lines, per local day

One row per (line, local day in Europe/Copenhagen). Ferries are public vessels.

Column	Type	Meaning
line	string	the ferry service, resolved at build time from OSM route endpoints
kind	string	island / domestic / international / foreign / harbour (see sql/40_ferry_trips.sql)
island	string	the island served, for kind = island; else empty
day	date	local date (Europe/Copenhagen) of the crossing's departure
year, season, dow	uint16, string, uint8	the baseline's grouping keys: calendar year, may-sep / oct-apr, ISO day of week
daytype	string	weekday / sat / sun, kept for coarse splits; not the baseline key
routes	uint64	how many OSM route objects were folded into the line that day
crossings	uint64	crossings that departed that local day, both directions
vessels	uint64	distinct vessels that made them
baseline	uint32	median crossings over the same line, year, season (May–Sep / Oct–Apr) and day of week, counting signal days only. Observed, not a timetable — Danish operator timetables are not archived per year
missed	int64	greatest(baseline - crossings, 0)
is_storm_day	uint8	1 if the local day is a named-storm date from DMI's list
fleet_positions	uint64	position messages heard from this line's fleet that day
fleet_sog_known	uint64	of those, with a usable speed field
fleet_moving	uint64	of those, moving
fleet_vessels_reporting	uint64	distinct vessels of the line heard at all

How to read a zero — the coverage columns exist for exactly this, and a consumer that prints "cancelled" without checking them is printing the receiver's bad day. Read the four cases in order and stop at the first that matches:

What you see	What it means
crossings = 0 and fleet_positions = 0	Silent. Nothing was heard from this line's fleet. Not a cancellation; not evidence.
fleet_positions > 0, fleet_sog_known = 0	Cannot tell. The vessel was heard, its speed field was empty all day, and neither a stay nor a crossing can be built from that.
fleet_moving = 0	The fleet lay still. A real cancellation, or a docking — unless crossings > 0, see the note on the fleet below.
fleet_moving > 0, crossings = 0	Moved, nothing matched. The matcher or OSM, not the weather.

The fleet_* columns count this line's own fleet, and a vessel-year belongs to exactly one line: the one on which it made the most crossings that year (own-majority, sql/41_ferry_daily.sql's rule; ties broken on the line name). So a relief vessel borrowed from another line has its crossings counted for the line it sailed, while its positions are counted for its home line — which is why 18 547 line-days in this file have fleet_moving = 0 and crossings > 0. Those are not still fleets; they are days worked by somebody else's boat.

Known biases

Read these before plotting anything.

1. msgs and moving_msgs are not comparable across September 2015. The archive duplicated its own feed 2015-08-28 → 2015-09-30: message counts in that window run ×2.199 on the Class A vessel-day median and ×2.552 on the daily mean. Flagged, never rewritten.
2. msgs and moving_msgs are not comparable across 2023 either. The Class A message tail doubles permanently between 2022-03 and 2023-11 while the median stands still: p90 8 749–9 428 → 16 261–19 607, p99 14 156–16 980 → 31 633–44 041. Vessel-days above the 43 200 position reports a Class A transponder can physically send in a day go from 0–0.06 % to 0.18–1.09 %; the worst day is 359 899 messages, 8.3× the ceiling. Any message count compared across 2023 compares two instruments.
3. vessels, and every head count, are comparable. Exact distinct counts are structurally immune to duplication — that is why head counts, not message counts, carry the cross-year story.
4. The loader drops rows, and says how many. Per loaded day: 4.0–11.3 % of rows as non-vessel (AtoN, base stations, SAR, PIRB, MOB), 0–6.3 % as outside the bounding box, 0.04–0.98 % as the Latitude = 91 sentinel.
5. Receiver reach barely moved — once. Of 139 sea regions, exactly one busy region steps on its head count: the Skagerrak approach at 57.9965 N 10.7602 E, 75 → 113 vessels (×1.507), between 2018-12 and 2021-01. The other eleven charted regions drift within ×0.82 .. ×1.16 over eleven years.
6. An independent source agrees on where the fleet is. Against EMODnet Human Activities' July-2021 leisure vessel density, Spearman ρ = 0.85 where both sources see the cell. Levels are not comparable (EMODnet measures track time, we measure presence: a moored boat is many vessel-hours to us and ~0 to EMODnet), so only rank was compared. 10.3 % of EMODnet's leisure hours fall where we have nothing, almost all of it outside the Danish core.

How to reproduce

git clone https://github.com/Maksim-Burtsev/seafolk
cd seafolk
# fetch and load the archive (see docs/PLAN.md; ~1.2 TB downloaded, deleted as
# it is aggregated, ~11 GB of ClickHouse storage left behind)
scripts/export.sh          # sql/70_export.sql -> dist/dataset/, then test_export.py

The export runs on clickhouse local --path data/ch (no server). The privacy test is part of scripts/export.sh, not an optional step.

Cite as

Burtsev, M. (2026). Seafolk — Danish AIS aggregates (version 0.1.0) [Data set].
Hugging Face. https://huggingface.co/datasets/mburtsev/seafolk-danish-ais
Source data: Danish Maritime Authority (aisdata.ais.dk).

@dataset{seafolk_danish_ais_2026,
  author    = {Burtsev, Maksim},
  title     = {Seafolk --- Danish AIS aggregates},
  year      = {2026},
  version   = {0.1.0},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/mburtsev/seafolk-danish-ais},
  note      = {Aggregated from the Danish Maritime Authority AIS archive,
               aisdata.ais.dk}
}

Versioning

The Hugging Face dataset, https://huggingface.co/datasets/mburtsev/seafolk-danish-ais, is the canonical address; a new version replaces the files there and bumps the version here. The code that builds every file is https://github.com/Maksim-Burtsev/seafolk. There is no DOI. Version here is 0.1.0.